JPH0632015B2 - Word detector - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a word detection device, and more particularly to a word detection device that detects a word included in an input voice in a voice recognition device, a voice input device, or the like.

(Prior Art) In a voice recognition device, a voice input device, etc., as one of the methods for detecting words in the input voice, the input voice is regarded as a sequence of categories such as syllables, phonemes, and phoneme classes, and detected from the input voice. There is a method in which if a category string created using the above categories corresponds to the category string of a word stored in the word dictionary, that word is used as the detection result.

In general, the above-mentioned categories have a short utterance time length, and there are similar categories. Therefore, it is difficult to detect only correct categories from the input speech without error. Therefore, conventionally, a method of detecting a plurality of category candidates for each category section in the input speech and then referring to the word dictionary to find a category candidate string corresponding to the word is used. However, even in this case, the presence of a certain category section may not be detected, or a correct category candidate may not be detected in a certain category section, due to the causes such as the looseness of vocalization and the adjacent categories such as articulation of syllables.

Therefore, Japanese Patent Application No. 61-190258, 190259, 1
As in the method described in “Word detection method” of 90260, 190261, after detecting the category candidate in the input speech, the category candidate is selected according to the category arrangement of the words in the word dictionary, and a certain category in the word is selected. When is not detected, the category candidates corresponding to the categories before and after that are used as a clue to find the category candidate sequence corresponding to the word category column, and the score for the word for which the corresponding category candidate sequence is found. Is taken to calculate.

On the other hand, when selecting a word with a good score from a large number of words, if the corresponding category candidate sequence and score are individually calculated for each word by the above method,
Especially when the number of words in the word dictionary is large, it requires a large amount of calculation and is not realistic.

Therefore, a tree-structured word dictionary that represents a category string of many words in a so-called tree structure is usually used. In such a word dictionary, the branches between the nodes correspond to the categories, and each branch row from the root node to the leaf node of the tree, that is, the category row, represents the category row of the word. Also, when multiple words begin with the same category string, the same category string is shared on the tree, which eliminates the need to traverse the same category string multiple times.

In addition, a score as a measure of the reliability in the detection evaluation is given to each of the category strings that have been traced from the root node to the middle node of the tree, and a category string with a good score, that is, a highly reliable category string is always selected and further advanced. By following, it is possible to finally obtain the word corresponding to the category sequence having a good score. This method is commonly referred to as the best-first search method.

According to this method, a word corresponding to a category string having a good score can be obtained without following the category string of all words. In this case, as the score of the category string, the average value of the scores of the category candidates in the category candidate string corresponding to the category string is always given.

(Problems to be Solved by the Invention) As described above, when the score of the category string is obtained from the average value of the scores of the category candidates in the corresponding category candidate string, the category close to the start end of the category string of the correct word When the score of the candidate corresponding to is significantly worse than the other candidates, there is a drawback that the word is hard to be detected.

FIG. 3 is an explanatory diagram showing an example of voice candidates extracted from the input voice. For example, suppose that syllables are used as categories as shown in FIG. 3, and when a voice "Zaysei" is input, a syllable candidate is obtained for each syllable. Here, the number written in each syllable candidate is the score of the syllable candidate, and the smaller the value, the better, that is, the more reliable. At this time, the syllable candidate sequence corresponding to each of the words “seigen (restriction)”, “seigen (finance source)”, “zeisei (tax system)”, and “seisei (financial)” is calculated using the syllable candidates shown in FIG. When created, the score is: Seigen: (3 + 7 + 8 + 4) /4=5.5 Seigen: (8 + 7 + 8 + 4) /4=6.75 Zeisei: (12 + 7 + 1 + 1) /4=5.25 Seisei: (8 + 7 + 1 + 1) /4=4.25 And the correct word "Zaysei" has the lowest score and therefore the highest score. However, the length from the beginning to the middle of the syllable string of each of these words is n.
When the score of the (n = 1,2,3,4) partial syllable string is obtained as the average value of the scores of the syllable candidates in the corresponding syllable candidate string as described above, the result is as follows.

Center: 3/1 = 3 Say: (3 + 7) / 2 = 5 Sayage: (3 + 7 + 8) / 3 = 6 Seigen: (3 + 7 + 8 + 4) /4=5.5 The: 8/1 = 8 Zay: (8 + 7) / 2 = 7.5 Zaige: (8 + 7 + 8) /3=7.66 Zeigen: (8 + 7 + 8 + 4) /4=6.75 Ze: 12/1 = 12 Zei: (12 + 7) /2=9.5 Zeise: (12 + 7 + 1) / 3 = 6.66 Zeise: (12 + 7 + 1 + 1) /4=5.25 The: 8/1 = 8 Zei: (8 + 7) /2=7.5 Zeise: (8 + 7 + 1) /3=5.33 Zeisei: (8 + 7 + 1 + 1) /4=4.25 Thus, the score of the syllable string consisting only of the leading syllable "Za" of the correct word "Zaisei" is not better than the score of any partial syllable string of the words "Seigen" and "Zaigen". . That is, the category sequence of the correct word “Zaisei” is traced only after the category sequences of the words “Saigen” and “Zaigen” are completed, which delays the detection of the correct word.

An object of the present invention is to eliminate the above-mentioned drawbacks and to execute a correct word faster and with less processing amount even when the score of a candidate corresponding to a category close to the beginning of the category sequence of a correct word is significantly worse than other candidates. An object of the present invention is to provide a word detection device that can be detected by.

(Means for Solving Problems) A word detection device of the present invention is a measure of reliability in detecting and evaluating a plurality of category candidates and category candidates from an input speech that is a sequence of categories such as syllables, phonemes, and phoneme classes. The category candidate extraction means for extracting and storing the position information of the score and the category sequence as a branch sequence from the root node to the leaf node as the category sequence of the detection target word A tree-structured word dictionary, a category string storage unit that stores at least one set of a category consisting of at least one category included in the word dictionary, and a category candidate sequence corresponding to the category, and the category string storage unit. When the number n of the category candidates in the corresponding category candidate sequence is equal to or more than the predetermined number N, the score of the category category Calculated using the average value of the scores, and when n is less than N, the number of virtual category candidates (n + m) is added to the number of category candidates of m as a preset function depending on n and N.
Score calculation means for calculating using the average value of the score of each category candidate, a category string with the best score among the category strings stored in the category string storage means, and a category candidate string corresponding to this If the category string with the highest score reaches the leaf node of the word dictionary, the word corresponding to that category string is output as the detection result, and if it does not reach the leaf node of the word dictionary, the category string And a category candidate sequence are output as undetected results, and the undetected results are received from the category candidate selection means, and the word dictionary is further traced from the end node of the category sequence to correspond to a longer category sequence. And a category string generating means for generating one or more sets of category candidate strings and adding them to the category string storing means. It is characterized in.

(Operation) In the above example, although the syllable candidate sequence corresponding to the syllable sequence “Zaysei” has a score of 4.25, which is better than the scores of other words, it corresponds to the leading syllable “The”. The score of the syllable candidate who does is bad as 8. On the other hand, the syllable candidate sequence corresponding to the syllable sequence “Saigen” has a score of 5.5, which is worse than that of the word “Zaisei”, but the score of the syllable candidate corresponding to the leading syllable “SE” is 3 and the word “Zaisei”. It is better than the score of the first syllable candidate.

In this way, the reliability of the word is correctly evaluated as a score calculated from the score of the entire category candidate string corresponding to the category string of the entire word, but in the prior art, the category corresponding to the category string of a part of the word Since the score calculated from only the candidate sequence is treated as the score of the word, the score of the word becomes worse when a part of the category accidentally contains a category candidate having a bad score.

On the other hand, when the category candidate string corresponding to only the category string of a part of the word is not determined, the score of the category candidate string corresponding to the category string of the entire word cannot be used.

Therefore, in the method of the present invention, among the categories in the word category string, the category candidate having a certain average score is hypothesized for the category for which the corresponding category candidate has not been determined yet. That is, when the score of a category candidate string is obtained, if the length n is shorter than a predetermined length N, it is determined that the category candidate string corresponds to a category string that is a part of a word. , N and N, the number m of category candidates is hypothesized. As a result, even when a part of the corresponding category candidate string includes a category candidate having a bad score, by averaging the score of the virtual category candidate string,
The word score is not so bad. Therefore, the category candidate sequence corresponding to the word can be quickly obtained.

For example, if N = 4 and m as a function preset depending on n and N is m = N−n, and if the score of a virtual syllable candidate is 1, the partial words of each word in the above example are represented. The scores for various example categories are as follows:

Center: (3 + 1 + 1 + 1) /4=1.5 Say: (3 + 7 + 1 + 1) / 4 = 3 Saege: (3 + 7 + 8 + 1) /4=4.75 Seigen: (3 + 7 + 8 + 4) /4=5.5 The: (8 + 1 + 1 + 1) / 4 = 2.75 Zai: (8 + 7 + 1 + 1) /4=4.25 Zaige: (8 + 7 + 8 + 1) / 4 = 6 Zeigen: (8 + 7 + 8 + 4) /4=6.75 Ze: (12 + 1 + 1 + 1) /4=3.75 Zei: (12 + 7 + 1 + 1) /4=5.25 Zeise: (12 + 7 + 1 + 1) /4=5.25 Zeise: (12 + 7 + 1 + 1) /4=5.25 The: (8 + 1 + 1 + 1) /4=2.75 Zei: (8 + 7 + 1 + 1) /4=4.25 Zeise: (8 + 7 + 1 + 1) /4=4.25 Zeise: (8 + 7 + 1 + 1) /4=4.25 Therefore, if you follow the syllable sequence in the order of this score, It will be traced in the following order. Here, the inside of "" is a corresponding word.

C: 1.5 “Saigen” The: 2.75 “Zaigen” The: 2.75 “Zaisei” Sei: 3 “Saigen” Zai: 4.25 “Zaigen” Zai: 4.25 “Zaisei” Saige: 4. 75 “Saigen” Seige: 6 “Zaigen” Zeise: 4.25 “Zaisei” Zaisei: 4.25 “Zaisei” Thus, the category sequence of the correct word “Zaisei” can be traced first. Also, the number of categories to follow is reduced.

(Example) Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. In the embodiment shown in FIG. 1, it is assumed that Japanese voice is input and syllables are used as categories.

The syllable candidate extraction unit 101 detects a syllable candidate in the input voice and stores the candidate in the syllable candidate storage unit 102 together with the score and the position in the input voice.

FIG. 2 is a block diagram showing an example of the syllable candidate extraction unit 101. In FIG. 2, the input voice is the voice buffer 20.
It is temporarily stored in 1. First, the vowel candidate detection unit 202 detects a vowel candidate in the voice stored in the voice buffer 201 and stores it in the vowel candidate storage unit 203. The vowel candidate is detected by comparing the voice standard pattern of each vowel stored in advance in the vowel pattern storage unit 204 with each section of the input voice. The vowel voice signal is relatively stationary and therefore easy to detect. Each vowel holds at least the vowel name and position information in the input voice.

After the detection of the vowel candidate is completed, the consonant candidate detection unit 205 detects the consonant candidate as described below. In Japanese, a syllable is a consonant (C) -vowel (V) pair. Therefore, in the input speech, a section within a certain time length (hereinafter, referred to as a VC section) of a section sandwiched between two vowels and a section within a certain time length from the start end of the input speech (hereinafter, referred to as a VC section). It can be said that there is one consonant in each of up to the CV section). The consonant candidate detection unit 204 stores the VCV section previously stored in the consonant pattern storage section 206 for all the VCV sections and CV sections created from the vowel candidates stored in the vowel candidate storage section 203.
And CV standard voice patterns are compared, and the names of a plurality of voice patterns having a high degree of similarity are used as consonant candidates. The vowel candidate and the consonant candidate determined in the above are combined into a syllable candidate, and the syllable candidate storage unit 102 is installed in the input voice.
Remember.

As an example, it is assumed that the voice "Zaisei (Finance)" is input. In this case, syllable candidates such as those shown in FIG. 3 are extracted as the syllable recognition result. In FIG. 3, a plurality of syllable candidates are extracted in each syllable section,
The number written in each syllable candidate is the score of that candidate.

The word dictionary 103 stores syllable strings of words to be detected in a tree structure format. Now, in the word dictionary 103, 7 words of "Keisan (calculation)", "Zaigen (finance source)", "Zaisan (property)", "Seigen (restriction)", "Sage (politics)", and "Zaysei (tax)" Is stored. This is shown in FIG. FIG. 4 is an explanatory diagram showing an example of the stored contents of the word dictionary in the embodiment of FIG. The numbers attached to the branches indicate branch numbers used in the following description.

First, the syllable candidate string generation unit 104 first determines the word dictionary 10
For each of the syllables of the branch following the root node of No. 3, when the syllable candidate corresponding to the syllable is stored in the syllable candidate storage unit 102, the syllable is regarded as a syllable string of length 1.
It is stored in the syllable string storage unit 105 together with the corresponding syllable candidate. Further, the score calculation unit 106 calculates and assigns the score to those syllable strings. In this embodiment, N
= 4, m = N−n, and the score of the virtual syllable candidate is 1.

For example, the syllable candidate corresponding to the branch center (13) is center [3], and the score of this syllable string is the average value of the scores of this syllable candidate, so that (3 + 1 + 1 + 1) /4=1.5. .

As a result, in this case, the syllable string storage unit 105 stores the following three syllable strings. Here, for each syllable string, the syllable string number, the score of the syllable string, the syllable string, and the corresponding syllable candidate string are shown in order. In addition, () is the branch number,
The value in [] is the score of the syllable candidate.

2.75 The (5) The [8] 1.5 C (13) C [3] 3.75 Z (18) Z [12] Next, the syllable string selection unit 107 stores the syllable string storage unit 105. Of the syllable strings, the syllable string having the highest score, that is, the smallest value is extracted, and the syllable string and the syllable candidate string are sent to the syllable string generation unit 104. The syllable string generation unit 104 further extends the received syllable string from its end point to generate a longer syllable string. That is, if the syllable candidate corresponding to the branch following the terminal node in the word dictionary 103 is included in the syllable candidate storage unit 102 and the syllable candidate can be connected to the current syllable candidate string, the syllable candidate is currently selected. Connect to the syllable candidate sequence of. The generated syllable string and syllable candidate string are stored in the syllable storage unit 105, and the score thereof is calculated by the score calculation unit 106.

In this case, the syllable string is extended. As a result, the syllable candidate sequence 3 (13) -A (14) C [3] -A [7] is stored in the syllable sequence storage unit 105. As a result, the syllable string storage unit stores the next syllable string.

2.75 the (5) the [8] 3.75 ze (18) ze [12] 3 c (13) -a (14) c [3] -a [7] Similarly, in the syllable string storage unit 105 The contents will be changed as follows.

A syllable string is generated from the syllable string.

 3.75 Ze (18) Ze [12] 3 Ce (13) -i (14) Ce [3] -i [7] 4.25 (5) -i (6) The [8] -i [7] A syllable string is generated from the syllable string.

 3.75 Ze (18) Ze [12] 4.25 The (5) -A (6) The [8] -A [7] 4.75 C (13) -A (14) -G (15) C [3) -a [7] -gear [8] A syllable string is generated from the syllable string.

 4.25 The (5) -a (6) The [8] -a [7] 4.75 The (13) -a (14) -ge (15) The [3] -a [7] -ge [ 8] 5.25 ze (18) -a (19) ze [12] -a [7] A syllable string is generated from the syllable string.

 4.75 Ce (13) -i (14) -Ge (15) Ce [3] -i [7] -Ge [8] 5.25 Ze (18) -i (19) Ze [12] -i [ 7] 6 the (5) -a (6) -ge (7) the [8] -a [7] -ge [8] 4.25 the (5) -a (6) -se (11) the [ 8] -a [7] -c [1] A syllable string is generated from the syllable string.

4.75 Ce (13) -i (14) -Ge (15) Ce [3] -i [7] -Ge [8] 5.25 Ze (18) -i (19) Ze [12] -i [ 7] 6 the (5) -a (6) -ge (7) the [8] -a [7] -ge [8] 4.25 the (5) -a (6) -se (11) -a (12) The [8] -a [7] -c [1] -a [1] Here, since the end of the syllable string reaches the leaf node of the word dictionary 103, the syllable string selecting unit 107 selects this word. "Zaysei" is output as the detection result.

Thus, the correct word "Zaysei" is first detected. In the present embodiment, in order to simplify the description, the case where no syllable recognition error occurs in the stage of syllable recognition, that is, the case where at least a correct syllable candidate is extracted for all input syllables is described. Even when a syllable recognition error occurs, the aforementioned Japanese Patent Application No. 61-190258,
By using the method described in "Word Detection Method" of 190259, 190260, 190261, a correct word can be detected efficiently as in the above embodiment.

(Effects of the Invention) As described above, according to the present invention, even when the score of the candidate corresponding to the category close to the start end of the category sequence of the correct word is significantly worse than the other candidates, the correct word is excluded. There is an effect that it is possible to realize a word detection device that can detect words earlier than, and can efficiently perform word detection with a small number of generated category strings.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a block diagram showing an example of a syllable candidate extraction unit in the embodiment of FIG. 1, and FIG. 3 is a syllable candidate extracted from an input voice. FIG. 4 is an explanatory diagram showing an example, and FIG. 4 is an explanatory diagram showing an example of stored contents of the word dictionary in the embodiment of FIG. 101 ... Syllable candidate extracting section, 102 ... syllable candidate storing section, 103 ... word dictionary, 104 ... syllable string generating section, 1
05 ... syllable string storage unit, 106 ... score calculation unit, 10
7 ... Syllable string selection unit, 201 ... voice buffer, 202
Vowel candidate detection unit, 203 vowel candidate storage unit, 20
4 ... Vowel pattern storage unit, 205 ... Consonant candidate detection unit,
206 ... Consonant pattern storage unit.

Claims (1)

[Claims] 1. A plurality of category candidates and a score and position information as a measure of reliability in detection and evaluation of these category candidates are extracted and stored from an input speech which is a sequence of categories such as syllables, phonemes, and phoneme classes. Category candidate extracting means, a tree-structured word dictionary in which branches between nodes correspond to categories, and a category string as a branch string from a root node to a leaf node is a category string of words to be detected, and the word dictionary A category string storing means for storing at least one set of a category string consisting of at least one category and a category candidate string corresponding to the category string, and a category corresponding to the score of the category string in the category string storing means. When the number n of category candidates in the candidate sequence is equal to or more than the predetermined number N, n is calculated using the average value of the category candidates, and n is When less than, it is calculated by using the average value of the scores of (n + m) category candidates obtained by adding m virtual category candidates as a preset function depending on n and N to n category candidates. A score calculating means and a category string having the highest score among the category strings stored in the category string storing means and a category candidate string corresponding thereto are extracted, and the category string having the highest score is a leaf node of a word dictionary. If the category string has not reached the leaf node of the word dictionary, the category string and the category candidate string are treated as undetected results. And a category selection means for outputting the result, and an undetected result is received from the category candidate selection means, and a word dictionary is further created from the node at the end of the category string. 1 set of a category string longer than the trace and a corresponding category candidate string
A word detecting device, comprising: a category string generating means for generating more than one and adding them to the category string storing means to detect a word.